JP7268367B2 - LEARNING DEVICE, LEARNING METHOD AND LEARNING PROGRAM - Google Patents

Description

The present invention relates to a learning device, a learning method, and a learning program.

In recent years, in the field of medical image analysis and the like, detection of abnormalities such as lesions using image data has been performed. Especially in the medical field, it is often difficult to obtain image data of abnormal conditions for use as training data. Since the data is handled, there is also a demand for faster detection speed.

Under these circumstances, as a technique for anomaly detection using machine learning, Generative Adversarial Networks (GAN) are used to estimate the distribution of normal data and detect data deviating from that distribution as abnormal data. It has been known. FIG. 17 is a diagram explaining the GAN. As shown in FIG. 17, GAN is a learning model of unsupervised learning that has a generator and a classifier, and a network that learns the generator and a network that learns the classifier grow together.

In GAN, the generator learns to generate fake data close to genuine data from input data such as noise, and the classifier learns to distinguish whether the data generated by the generator is genuine data or fake data. . As anomaly detection using this GAN, there is a method of determining whether or not a learned generator has the ability to generate a given sample, and if it does not have the ability, it is regarded as abnormal data, or a method of learning identification A technique is used in which data determined by the device to be fake data is regarded as abnormal data.

Schlegl, Thomas, et al. “Unsupervised anomaly detection with generative adversarial networks to guide marker discovery.”, International Conference on Information Processing in Medical Imaging. Springer, Cham, 2017. M. Sabokrou, et al. “Adversarially Learned One-Class Classifier for Novelty Detection,” Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. 2018.

However, with the above technique, the ability of the discriminator to detect anomalies may be degraded. Specifically, when a discriminator is used, the characteristics of the discriminator depend on data other than normal data that is erroneously generated by the generator at the end of learning, which may result in large detection omissions. For example, a discriminator can be expected to be capable of discerning the difference between normal data and the generator's output, while anything that is not in the generator's output and not in the normal data cannot be learned because , do not expect any specific output. It should be noted that the method using the generator is not suitable for anomaly detection that requires high speed because the detection cost is high.

An object of one aspect of the present invention is to provide a learning device, a learning method, and a learning program capable of suppressing deterioration in the anomaly detection capability of a discriminator.

In the first scheme, the learning device has a generator that generates image data and a discriminator that discriminates whether the image data is true or false. The learning device has a learning unit that performs learning so that the generator maximizes the discrimination error of the classifier and the classifier learns so that the classifier minimizes the discrimination error. The learning device has a storage unit that stores the state of the generator during learning under preset conditions during execution of the learning. The learning device has a re-learning unit that re-learns the classifier using the saved state of the classifier during learning.

According to one embodiment, it is possible to suppress deterioration of the abnormality detection capability of the discriminator.

FIG. 1 is a diagram for explaining an abnormality detection device according to a first embodiment. FIG. 2 is a diagram for explaining before learning. FIG. 3 is a diagram for explaining the state after learning has started. FIG. 4 is a diagram for explaining the progress of learning. FIG. 5 is a diagram for explaining a discriminator after learning. FIG. 6 is a functional block diagram of the functional configuration of the abnormality detection device according to the first embodiment; FIG. 7 is a diagram illustrating an example of detection results. FIG. 8 is a flowchart illustrating the flow of learning processing according to the first embodiment; FIG. 9 is a flowchart illustrating the flow of detection processing according to the first embodiment; FIG. 10 is a diagram for explaining temporary storage and relearning. FIG. 11 is a diagram for explaining the effect. FIG. 12 is a diagram for explaining learning processing according to the second embodiment. FIG. 13 is a flowchart illustrating the flow of learning processing according to the second embodiment. FIG. 14 is a diagram for explaining learning processing according to the third embodiment. FIG. 15 is a flowchart illustrating the flow of learning processing according to the third embodiment. FIG. 16 is a diagram illustrating a hardware configuration example. FIG. 17 is a diagram explaining the GAN.

Embodiments of the learning device, the learning method, and the learning program disclosed in the present application will be described in detail below with reference to the drawings. In addition, this invention is not limited by this Example. Moreover, each embodiment can be appropriately combined within a range without contradiction.

[Description of the abnormality detection device 10]
FIG. 1 is a diagram for explaining an abnormality detection device 10 according to a first embodiment. As shown in FIG. 1 , in the learning phase, the anomaly detection apparatus 10 learns a generator and a discriminator using a GAN, and in the prediction phase, uses the learned discriminator to predict image data. It is an example of a learning device that performs anomaly detection.

By the way, in general GAN-based learning, the ability of the discriminator to detect anomalies may deteriorate. Here, problems of general GANs will be described with reference to FIGS. 2 to 5. FIG. Here, FIG. 2 to FIG. 5 illustrate and explain each stage from before the start of learning to after the end of learning with respect to learning by general GAN. FIG. 2 is a diagram for explaining before learning. FIG. 3 is a diagram for explaining the state after learning has started. FIG. 4 is a diagram for explaining the progress of learning. FIG. 5 is a diagram for explaining a discriminator after learning.

As shown in FIG. 2, before learning, the generator generates data with a distribution different from normal data, and the discriminator learns the difference between the distribution of normal data and the distribution of data generated by the generator. . That is, the discriminator clearly distinguishes and discriminates the distribution of normal data and the distribution of data generated by the generator.

Subsequently, as shown in FIG. 3, when learning is started, the learning of the generator progresses so that more data is identified as normal by the classifier and less data is identified as abnormal. That is, the generator learns to generate data that causes the discriminator to misidentify, so that the distribution of the data generated by the generator is learned to include a range in which the discriminator can determine that it is normal. On the other hand, when the distribution of data generated by the generator changes with the start of learning, the discriminator changes the probability of discrimination results to be normal. Learn the difference from the distribution of the data generated by the instrument.

3, the distribution of data generated by the generator is learned so as to include a range in which the discriminator can identify normal in FIG. 3, as shown in FIG. That is, the distribution of data generated by the generator is updated so that it includes more of the range that the classifier can identify as normal. On the one hand, the discriminator learns such that the updated generator discriminates between the distribution of generated data and the distribution of normal data.

As learning progresses further, as shown in FIG. 5, the generator learns from the state of FIG. is updated so as to include the range that can be discriminated by the discriminator in FIG. When learning is completed, the discriminator after learning is trained so that it can discriminate between data generated by the generator and normal data (genuine data).

However, in a general GAN, since the characteristics of the discriminator depend on data other than normal data erroneously generated by the generator at the end of learning, detection omissions may increase. In other words, an abnormality can be correctly detected for abnormal data existing in a range in which the discriminator after learning can be determined to be abnormal, but abnormal data existing in other ranges cannot be detected. In other words, in general GANs, it may not be possible to learn classifiers up to the range of anomalous data that is expected to be detected.

Therefore, the anomaly detection apparatus 10 according to the first embodiment focuses on the fact that anomalous data to be detected in medical care does not appear in a location greatly deviating from the distribution of normal data. Since the distribution of data generated by the generators may temporarily deviate from the normal data distribution during the learning process, the abnormality detection apparatus 10 stores the generators in the process of learning at regular intervals. and train a discriminator using all the stored generators.

Specifically, as shown in FIG. 1, the anomaly detection apparatus 10 learns generators and discriminators based on the GAN, and acquires information about generators in the process of learning at regular intervals. Then, when learning of the discriminator by GAN is completed, the anomaly detection apparatus 10 re-learns the learned discriminator using each generator saved in the middle. After completing the re-learning, the anomaly detection device 10 inputs the prediction target image data to the re-learned classifier, and performs anomaly detection.

[Function configuration]
FIG. 6 is a functional block diagram showing the functional configuration of the abnormality detection device 10 according to the first embodiment. As shown in FIG. 6 , the abnormality detection device 10 has a communication section 11 , a storage section 12 and a control section 20 .

The communication unit 11 is a processing unit that controls communication between other devices, such as a communication interface. For example, the communication unit 11 receives various processing start instructions, normal data, prediction target data, and the like from the administrator terminal, and transmits learning results, prediction results, and the like to the administrator terminal.

The storage unit 12 is an example of a storage device that stores various data, programs executed by the control unit 20, and the like, and is, for example, a memory or a hard disk. The storage unit 12 stores a normal data DB 13, an intermediate information DB 14, a learning result DB 15, and the like.

The normal data DB 13 is a database that stores image data obtained by imaging internal organs in a normal state. For example, the normal data DB 13 stores normal data used in learning of classifiers by GAN.

The in-progress information DB 14 is a database that stores information about generators in the process of learning in the process of learning generators by GAN. For example, the in-progress information DB 14 stores various parameters in the process of learning that can reproduce (construct) the generator in the process of learning.

The learning result DB 15 is a database that stores learning results of generators and classifiers. For example, the learning result DB 15 stores the learning results of generators and discriminators using GAN, and various parameters that can generate learned generators and discriminators for which learning has been completed as re-learning results of discriminators.

The control unit 20 is a processing unit that controls the entire abnormality detection device 10, such as a processor. The control unit 20 has a learning processing unit 30, a re-learning unit 40, and a detection unit 50, learns a classifier with high recognition accuracy, and executes anomaly detection on input prediction target image data.

The learning processing unit 30 has a generator learning unit 31, a discriminator learning unit 32, and a temporary storage unit 33, and is a processing unit that executes learning of the generator and discriminator using GAN.

The generator learning unit 31 is a processing unit that executes GAN generator learning. Specifically, the generator learning unit 31 learns the generator so that the discriminator has more data determined to be normal and less data determined to be abnormal, as in a general GAN. That is, the generator learns to maximize the discrimination error of the discriminator.

For example, the generator learning unit 31 generates image data using a generator using a latent variable such as a seed, which is a random number or random noise. Then, the generator learning unit 31 inputs the generated image data (hereinafter sometimes referred to as generated data) to the classifier, and acquires the classification result of the classifier. Then, the generator learning unit 31 uses the identification result to learn the generator so that the generated data is identified as normal data by the identifier.

That is, the generator learning unit 31 learns the generator so that the distribution of generated data generated by the generator matches the distribution of normal data. When the learning is completed, the generator learning unit 31 stores information on the learned generator in the learning result DB 15 .

The discriminator learning unit 32 is a processing unit that performs learning of a GAN discriminator. Specifically, the discriminator learning unit 32 learns a generator so as to distinguish between normal data stored in the normal data DB 13 and generated data generated by the generator, as in a general GAN. That is, the discriminator learns to minimize the discrimination error.

For example, the discriminator learning unit 32 acquires generated data from the generator, inputs it to the discriminator, and acquires the output result of the discriminator. Then, the discriminator learning unit 32 learns the discriminator so that the normality probability of the normal data included in the output result and the abnormality probability of the generated data are increased and the normality probability is decreased. do.

That is, the discriminator learning unit 32 causes the discriminator to learn the difference between the distribution of generated data generated by the generator and the distribution of normal data. When the learning is completed, the discriminator learning unit 32 stores information about the discriminators that have been trained in the learning result DB 15 .

The temporary storage unit 33 is a processing unit that stores the generator in the middle of learning. For example, in the learning process of the generator by the generator learning unit 31, the temporary storage unit 33 acquires various parameters indicating the state of the generator during learning from the generator at regular time intervals or at regular learning times. and store it in the intermediate information DB 14.

The re-learning unit 40 is a processing unit that re-learns the discriminator learned by the GAN. Specifically, the re-learning unit 40 re-learns classifiers using all of the generators in the middle of learning that are stored in the temporary storage unit 33 .

For example, when the learning by the learning processing unit 30 is completed, the relearning unit 40 reads information such as parameters from the in-progress information DB 14 and generates each generator that is in the process of learning. Similarly, the re-learning unit 40 reads classifier parameters and the like from the learning result DB 15 and generates a trained classifier.

Here, it is assumed that four discriminators in the middle of learning are generated. Then, the re-learning unit 40 generates generated data using the first generator, inputs the generated data to a learned classifier, and causes the classifier to learn. Subsequently, the re-learning unit 40 constructs generated data using the second generator, inputs it to the classifier trained by one classifier, and trains the classifier. In this manner, the re-learning unit 40 re-learns the discriminators by sequentially using each generator in the middle of learning. Then, when the re-learning is completed, the re-learning unit 40 stores information about the re-learned classifier in the learning result DB 15 .

The detection unit 50 is a processing unit that performs abnormality detection from input image data. For example, when the re-learning by the re-learning unit 40 is completed, the detection unit 50 reads out the parameters related to the re-learned classifier from the learning result DB 15 and builds the classifier. Then, when receiving the image data to be predicted, the detection unit 50 inputs the image data to the constructed discriminator, and obtains an output result.

Here, the detection unit 50 regards the input image data as normal image data when the normality probability is higher than the normality probability of the normal data included in the output result and the abnormality probability of the generated data. If the abnormality probability is higher, the input image data is identified as abnormal image data. Then, the detection unit 50 transmits the identification result to the administrator terminal, displays it on a display unit such as a display, or stores it in the storage unit 12 .

FIG. 7 is a diagram illustrating an example of detection results. As shown in FIG. 7, the detection unit 50 divides the input image to be predicted into small regions, and then inputs the divided regions to the classifier learned by the re-learning unit 40 to acquire the classification result. Then, the detection unit 50 can present to the user the region determined to be abnormal by the discriminator.

[Flow of learning process]
FIG. 8 is a flowchart illustrating the flow of learning processing according to the first embodiment; As shown in FIG. 8, when learning processing is instructed, the learning processing unit 30 initializes the generator and the discriminator (S101).

Subsequently, the generator learning unit 31 generates data (generated data) in the generator (S102), and the classifier learning unit 32 learns the classifier so that normal data and generated data can be distinguished (S103). ). Then, the generator learning unit 31 learns the generator so that the generated data is identified as normal data by the classifier (S104).

After that, when the generator is not stored for a certain amount of time or the number of iterations (number of times of learning) is equal to or greater than a certain value (S105: Yes), the temporary storage unit 33 stores the generator in the process of learning (S106). If the generator is not saved for a period of time or the number of iterations is less than a certain value (S105: No), S107 is executed without executing S106.

Thereafter, S102 and subsequent steps are repeated until the learning ends (S107: No). On the other hand, when the learning ends (S107: Yes), the re-learning unit 40 re-learns the discriminator so that it can discriminate between the normal data and the data generated by the stored generator (generated data) ( S108). After completing the re-learning, the re-learning unit 40 stores the classifier (S109).

[Flow of detection process]
FIG. 9 is a flowchart illustrating the flow of detection processing according to the first embodiment; As shown in FIG. 9, when the detection process is instructed, the detection unit 50 reads the saved discriminator (S201). Subsequently, the detection unit 50 inputs the target data to the discriminator and determines (identifies) an abnormality of the target data (S202).

Here, the detection unit 50 ends the process when it is determined to be normal (S203: Yes), and outputs the target data as abnormal data when it is not determined to be normal (S203: No) (S204).

[effect]
As described above, the anomaly detection apparatus 10 focuses on the fact that anomalous data to be detected, for example, in medical care does not appear in a place that deviates greatly from the distribution of normal data, and re-learns the classifier using a generator that is in the process of learning. to run. As a result, the abnormality detection apparatus 10 can re-learn the classifier using a generator capable of comprehensively generating abnormal data expected to exist around normal data. can be extended.

FIG. 10 is a diagram for explaining temporary storage and relearning. Saving a generator that is in the process of learning means saving the state of the generator that is in the process of learning. will be saved. That is, as shown in FIG. 10, the learning process of the distribution of each generated data generated by the generator can be saved from the beginning of learning to the completion of learning. Therefore, by re-learning using all the stored generators, it is possible to learn each distribution of the generated data of each generator, so that the range that the classifier that has completed learning by GAN cannot cover can learn.

FIG. 11 is a diagram for explaining the effect. As shown in FIG. 11, in a general GAN, the characteristics of the discriminator depend on data other than normal data erroneously generated by the generator at the end of learning. It is difficult to learn a discriminator immediately, and there are many detection omissions. In the example of FIG. 11, the discriminator cannot detect data A outside the data distribution range of the generator as abnormal.

On the other hand, in the learning of the discriminator according to the first embodiment, the distribution of the generated data generated by the generator that is saved in the middle can be re-learned. , can cover more of the periphery of normal data. As a result, the discriminator after re-learning can detect data A, which could not be detected by a discriminator based on a general GAN, as abnormal data.

By the way, in the first embodiment, an example in which generators in the middle of learning are stored at regular intervals has been described. Therefore, in the second embodiment, an example will be described in which generators are stored not at regular intervals but at timings when the loss of the discriminator temporarily rises above the threshold and then falls below the threshold again.

FIG. 12 is a diagram for explaining learning processing according to the second embodiment. FIG. 12 illustrates the time change of the loss of the discriminator, that is, the time change of the rate of failure of the discriminator in learning by GAN. As shown in FIG. 12, as the learning of the generator by the GAN progresses, the distribution of the data generated by the generator begins to deviate to the area not covered by the discriminator ((a) in FIG. 12), and the discriminator loss increases (FIG. 12(b)). After that, the learning of the discriminator progresses and the discriminator follows the learning of the generator, so the loss of the discriminator decreases ((c) in FIG. 12) and the deviation is eliminated (((c) in FIG. 12). d)).

Therefore, the temporary storage unit 33 stores the state of the generator when the loss of the discriminator increases and the loss starts to decrease. If it is desired to store a plurality of generators at this timing, the temporary storage unit 33 can also store the generators at regular time intervals after the loss of the discriminator begins to decrease. By doing so, it is possible to store the point in time when there is a large difference between the distribution of data generated by the generator and the distribution of normal data, so that the learning range of the discriminator can be widened. As a result, the performance of the discriminator can be improved.

FIG. 13 is a flowchart illustrating the flow of learning processing according to the second embodiment. As shown in FIG. 13, when learning processing is instructed, the learning processing unit 30 initializes the generator and the discriminator (S301).

Subsequently, the generator learning unit 31 generates data (generated data) in the generator (S302), and the classifier learning unit 32 learns the classifier so that normal data and generated data can be distinguished (S303). ). Then, the generator learning unit 31 learns the generator so that the generated data is identified as normal data by the classifier (S304).

After that, the temporary storage unit 33 is in the process of learning when the generator is not stored for a certain amount of time or the number of iterations is greater than a certain value (S305: Yes) and the loss of the discriminator is decreasing (S306: Yes). Save the generator (S307). If the length of time the generator is not stored or the number of iterations is less than a certain value (S305: No), or if the loss of the discriminator has not decreased (S306: No), S308 is executed without executing S307. be done.

Thereafter, S302 and subsequent steps are repeated until the learning ends (S308: No). On the other hand, when the learning ends (S308: Yes), the re-learning unit 40 re-learns the discriminator so that it can discriminate between the normal data and the data generated by the stored generator (generated data) ( S309). After completing the re-learning, the re-learning unit 40 stores the discriminator (S310). Note that the detection process is the same as that of the first embodiment, so a detailed description thereof will be omitted.

By the way, in the second embodiment, when the generator starts generating abnormal data, the discriminator immediately follows. Also, it may be difficult to determine how far the deviation has progressed.

Therefore, in the third embodiment, two discriminators with different learning speeds are used, and the generator in the advanced deviation state is accurately stored, thereby improving the performance of the discriminator. Specifically, the generator learns using a slow discriminator and measures the progress of deviation using a fast discriminator. In the second embodiment, the ratio of generated data that can be identified as generated data by a fast discriminator is used as an index. In addition, when the deviation progresses beyond a certain level, the generator is saved, and the fast generator is copied to the slow discriminator so that no further deviation occurs.

FIG. 14 is a diagram for explaining learning processing according to the third embodiment. As shown in FIG. 14, in FIG. 14, a discriminator A with a slow learning speed (hereinafter, may be referred to as a discriminator A, a slow discriminator A, etc.) and a discriminator B with a fast learning speed (hereinafter referred to as In learning by GAN using two discriminators B, fast discriminator B, etc.), the time change of the loss of each discriminator is illustrated.

As shown in FIG. 14, as the learning of the generator by the GAN progresses, the distribution of the data generated by the generator begins to deviate to the area not covered by the fast discriminator B ((a) in FIG. 14), The loss of fast discriminator B increases (FIG. 14(b)). However, since the slow discriminator A has a slow learning speed, the increase in loss is smaller than that of the fast discriminator B.

After that, the learning of both classifiers progresses, but the learning of the particularly fast classifier B progresses and follows the learning of the generator, so the loss of the fast classifier B decreases, and the slow classifier A allows deviation. ((c) in FIG. 14). That is, the slow discriminator A cannot follow the learning of the generator, and correction of discrimination does not progress.

At this time, the temporary storage unit 33 stores the generator in the middle of learning. That is, the temporary storage unit 33 stores a generator that learns using the slow classifier A when the deviation of the fast classifier B has progressed beyond a certain level. In other words, the temporary storage unit 33 stores the state of the generator when the loss of the slow discriminator B is increasing while the loss of the fast discriminator A starts to decrease after increasing. Note that a common value can be used for the discriminator A and the discriminator B as the threshold for determining whether the loss is equal to or greater than a certain value.

Then, the temporary storage unit 33 copies the parameters such as the weight of the fast discriminator to the slow discriminator A, and resumes learning using the two learning speeds ((d) in FIG. 14). That is, both of the two discriminators are changed to the fast discriminator state, and slow learning and fast learning are performed from that state.

When the learning using the two discriminators is completed in this way, the re-learning unit 40 re-learns the discriminator B using the temporarily stored generator in the middle of learning.

FIG. 15 is a flowchart illustrating the flow of learning processing according to the third embodiment. As shown in FIG. 15, when the learning process is instructed, the learning processing unit 30 initializes the generator, classifier A, and classifier B (S401).

Subsequently, the generator learning unit 31 sets the learning speed of the discriminator A to be lower than that of the discriminator B (S402). For example, the generator learning unit 31 slows down the learning speed by lowering the learning rate or lowering the learning frequency.

Then, the generator learning unit 31 generates data (generated data) in the generator (S403), and the discriminator learning unit 32 uses the discriminator A and the discriminator B so that normal data and generated data can be discriminated. Learn (S404). Then, the generator learning unit 31 learns the generator so that the generated data is identified as normal data by the classifier A (S405).

After that, if the generator is not stored for a certain amount of time or the number of iterations is more than a certain amount (S406: Yes), and the loss of classifier B is decreasing (S407: Yes), the temporary storage unit 33 A certain generator is saved (S408). Subsequently, the temporary storage unit 33 copies the parameters such as the weight of the discriminator B to the discriminator A (S409).

Note that if the generator is not stored for a period of time or the number of iterations is less than a certain value (S406: No), or if the loss of the discriminator has not decreased (S407: No), S410 is performed without executing S408 and S409. is executed.

Thereafter, S403 and subsequent steps are repeated until the learning ends (S410: No). On the other hand, when learning ends (S410: Yes), the re-learning unit 40 re-learns the discriminator B so that it can discriminate between normal data and data generated by the stored generator (generated data). (S411). After completing the re-learning, the re-learning unit 40 stores the discriminator B (S412). Note that the detection process is the same as that of the first embodiment, so a detailed description thereof will be omitted.

Although the embodiments of the present invention have been described so far, the present invention may be implemented in various different forms other than the embodiments described above.

[study]
For example, the timing of ending the learning process can be arbitrarily set, such as when learning using a predetermined number or more of learning data is completed, or when the loss of the discriminator becomes less than a threshold. In addition, the present invention can be applied not only to medical image data, but also to various other fields such as illegal intrusion and bringing in of dangerous substances.

[Timing of re-learning]
In the above embodiment, an example of re-learning the discriminator after learning is completed has been described, but the present invention is not limited to this. For example, the discriminator can concurrently perform learning using a normal GAN and learning using a generator that is temporarily stored and is in the process of learning. In other words, in order to suppress the storage of a generator that regenerates abnormal data that can already be generated by the stored generator, the stored generator is used for learning of the discriminator even during learning. By doing so, the learning time can be shortened, and the learning of the discriminator can be corrected while learning, so that the discriminating performance of the discriminator can be improved.

[system]
Information including processing procedures, control procedures, specific names, and various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified.

Also, each component of each device illustrated is functionally conceptual, and does not necessarily need to be physically configured as illustrated. That is, the specific forms of distribution and integration of each device are not limited to those shown in the drawings. That is, all or part of them can be functionally or physically distributed and integrated in arbitrary units according to various loads and usage conditions. For example, the learning processing unit 30, the re-learning unit 40, and the detection unit 50 can be realized by separate devices.

Further, each processing function performed by each device may be implemented in whole or in part by a CPU and a program analyzed and executed by the CPU, or implemented as hardware based on wired logic.

[hardware]
FIG. 16 is a diagram illustrating a hardware configuration example. As shown in FIG. 16, the abnormality detection device 10 has a communication device 10a, a HDD (Hard Disk Drive) 10b, a memory 10c, and a processor 10d. 16 are interconnected by a bus or the like.

The communication device 10a is a network interface card or the like, and communicates with other servers. The HDD 10b stores programs and DBs for operating the functions shown in FIG.

The processor 10d reads from the HDD 10b or the like a program for executing the same processing as each processing unit shown in FIG. 6 and develops it in the memory 10c, thereby operating processes for executing each function described with reference to FIG. 6 and the like. That is, this process executes the same function as each processing unit of the abnormality detection device 10 . Specifically, the processor 10d reads from the HDD 10b or the like a program having the same functions as those of the learning processing section 30, the re-learning section 40, the detection section 50, and the like. Then, the processor 10d executes processes for executing the same processes as the learning processing unit 30, the re-learning unit 40, the detecting unit 50, and the like.

Thus, the abnormality detection device 10 operates as an information processing device that executes an abnormality detection method by reading and executing a program. Further, the abnormality detection device 10 can read the program from the recording medium by the medium reading device and execute the read program, thereby realizing the same function as the embodiment described above. Note that the program referred to in this other embodiment is not limited to being executed by the abnormality detection device 10 . For example, the present invention can be applied in the same way when another computer or server executes the program, or when they cooperate to execute the program.

10 Abnormality Detection Device 11 Communication Unit 12 Storage Unit 13 Normal Data DB
14 Intermediate information DB
15 Learning result database
20 control unit 30 learning processing unit 31 generator learning unit 32 discriminator learning unit 33 temporary storage unit 40 re-learning unit 50 detection unit

Claims (8)

a generator for generating image data;
a discriminator that discriminates between genuine image data and generated image data ;
In the learning process, the generator learns to maximize the discrimination error of the discriminator, and using data generated by the generator, the discriminator learns to minimize the discrimination error. a learning unit that performs learning processing ;
The discriminator trained with the data generated by the generator in the first state using the data generated by the generator in the second state, which is the state prior to the first state in the learning process. and a relearning unit that performs relearning of the learning device. The second state is a state before the first state, and is a state when the loss of the discriminator becomes less than the threshold again after becoming equal to or greater than the threshold. A learning device according to claim 1. The discriminator includes a first discriminator and a second discriminator having a faster learning speed than the first discriminator,
In the learning process, the generator learns to maximize the identification error of the first classifier, and the first classifier learns the first classifier using data generated by the generator. A learning process in which the classifier learns to minimize the identification error and the second classifier learns to minimize the identification error of the second classifier,
The second state is a state prior to the first state, in which the loss of the second discriminator starts to decrease after increasing, and when the loss of the first discriminator is equal to or greater than a threshold in the state of
The relearning unit uses the data generated by the generator in the second state to perform relearning of the second discriminator trained with the data generated by the generator in the first state. 2. The learning device according to claim 1, wherein: 2. The learning device according to claim 1 , wherein the relearning unit starts the relearning after the learning process by the learning unit is completed . 2. The learning device according to claim 1, wherein the relearning unit starts the relearning before the learning process by the learning unit is completed . When image data to be classified is divided into a plurality of regions and input to the relearned classifier that has been relearned by the relearning unit , and an abnormality is detected by the relearned classifier 2. The learning device according to claim 1, further comprising a detection unit that presents an area indicated as abnormal by the re-learned classifier. In the learning process, a generator that generates image data learns to maximize the identification error of a classifier that identifies real image data and generated image data , and the data generated by the generator is learned , performing a learning process in which the discriminator learns to minimize the discriminant error,
The discriminator trained with the data generated by the generator in the first state using the data generated by the generator in the second state, which is the state prior to the first state in the learning process. A learning method characterized in that a computer executes processing for re-learning. In the learning process, a generator that generates image data learns to maximize the identification error of a classifier that identifies real image data and generated image data , and the data generated by the generator is learned , performing a learning process in which the discriminator learns to minimize the discriminant error,
The discriminator trained with the data generated by the generator in the first state using the data generated by the generator in the second state, which is the state prior to the first state in the learning process. A learning program characterized by causing a computer to execute a process of re-learning.

Images